Seventeen dated, verified operational quirks across two venues — fee conventions, funding grids, decimal scaling, accounting identities, API churn — each paired with a machine-actionable agent rule. Every entry was learned from live measurement, not documentation.
§1Capabilities
Captures the operational reality documentation misses: that fees are tiered per-symbol data rather than a constant, that prices, quantities, and collateral use three different decimal scalings, that the venue publishes its own config-drift signals on-chain.
Pairs every quirk with an agent rule — a structured directive an automated system can load and obey, not a paragraph a human has to remember.
Versions and dates everything, so when a venue changes behavior you can diff what your agent believed against what is now true.
§2Design
Schema-first. Every entry validates against a published JSON schema in CI. A malformed quirk cannot ship.
Measured, not transcribed. Entries come from live probes, drift sentinels, and chain reads — with the measurement date on the record.
Feeds the tools. The MCP server and the SDK read the registry directly. One correction propagates to every agent consuming it.
§3Availability by tier
| Open | Point-in-time snapshot, CC-BY. |
| Pro | Continuously updated multi-venue feed from our drift monitoring. |
| Institutional | Custom venue coverage on request. |
Indicative until launch.
§4 · Stated plainly
Snapshot covers two venues today. The registry is only as current as its dates — which is exactly why the dates are on every entry.